crypto/sha/asm/sha512-armv8.pl

   1 #!/usr/bin/env perl
   2 #
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
   5 # project. The module is, however, dual licensed under OpenSSL and
   6 # CRYPTOGAMS licenses depending on where you obtain it. For further
   7 # details see http://www.openssl.org/~appro/cryptogams/.
   8 # ====================================================================
   9 #
  10 # SHA256/512 for ARMv8.
  11 #
  12 # Performance in cycles per processed byte and improvement coefficient
  13 # over code generated with "default" compiler:
  14 #
  15 #               SHA256-hw       SHA256(*)       SHA512
  16 # Apple A7      1.97            10.5 (+33%)     6.73 (-1%(**))
  17 # Cortex-A53    2.38            15.5 (+115%)    10.0 (+150%(***))
  18 # Cortex-A57    2.31            11.6 (+86%)     7.51 (+260%(***))
  19 #
  20 # (*)   Software SHA256 results are of lesser relevance, presented
  21 #       mostly for informational purposes.
  22 # (**)  The result is a trade-off: it's possible to improve it by
  23 #       10% (or by 1 cycle per round), but at the cost of 20% loss
  24 #       on Cortex-A53 (or by 4 cycles per round).
  25 # (***) Super-impressive coefficients over gcc-generated code are
  26 #       indication of some compiler "pathology", most notably code
  27 #       generated with -mgeneral-regs-only is significanty faster
  28 #       and the gap is only 40-90%.
  29
  30 $flavour=shift;
  31 $output=shift;
  32
  33 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  34 ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
  35 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
  36 die "can't locate arm-xlate.pl";
  37
  38 open OUT,"| \"$^X\" $xlate $flavour $output";
  39 *STDOUT=*OUT;
  40
  41 if ($output =~ /512/) {
  42         $BITS=512;
  43         $SZ=8;
  44         @Sigma0=(28,34,39);
  45         @Sigma1=(14,18,41);
  46         @sigma0=(1,  8, 7);
  47         @sigma1=(19,61, 6);
  48         $rounds=80;
  49         $reg_t="x";
  50 } else {
  51         $BITS=256;
  52         $SZ=4;
  53         @Sigma0=( 2,13,22);
  54         @Sigma1=( 6,11,25);
  55         @sigma0=( 7,18, 3);
  56         @sigma1=(17,19,10);
  57         $rounds=64;
  58         $reg_t="w";
  59 }
  60
  61 $func="sha${BITS}_block_data_order";
  62
  63 ($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
  64
  65 @X=map("$reg_t$_",(3..15,0..2));
  66 @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
  67 ($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
  68
  69 sub BODY_00_xx {
  70 my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
  71 my $j=($i+1)&15;
  72 my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
  73    $T0=@X[$i+3] if ($i<11);
  74
  75 $code.=<<___    if ($i<16);
  76 #ifndef __ARMEB__
  77         rev     @X[$i],@X[$i]                   // $i
  78 #endif
  79 ___
  80 $code.=<<___    if ($i<13 && ($i&1));
  81         ldp     @X[$i+1],@X[$i+2],[$inp],#2*$SZ
  82 ___
  83 $code.=<<___    if ($i==13);
  84         ldp     @X[14],@X[15],[$inp]
  85 ___
  86 $code.=<<___    if ($i>=14);
  87         ldr     @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
  88 ___
  89 $code.=<<___    if ($i>0 && $i<16);
  90         add     $a,$a,$t1                       // h+=Sigma0(a)
  91 ___
  92 $code.=<<___    if ($i>=11);
  93         str     @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
  94 ___
  95 # While ARMv8 specifies merged rotate-n-logical operation such as
  96 # 'eor x,y,z,ror#n', it was found to negatively affect performance
  97 # on Apple A7. The reason seems to be that it requires even 'y' to
  98 # be available earlier. This means that such merged instruction is
  99 # not necessarily best choice on critical path... On the other hand
 100 # Cortex-A5x handles merged instructions much better than disjoint
 101 # rotate and logical... See (**) footnote above.
 102 $code.=<<___    if ($i<15);
 103         ror     $t0,$e,#$Sigma1[0]
 104         add     $h,$h,$t2                       // h+=K[i]
 105         eor     $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
 106         and     $t1,$f,$e
 107         bic     $t2,$g,$e
 108         add     $h,$h,@X[$i&15]                 // h+=X[i]
 109         orr     $t1,$t1,$t2                     // Ch(e,f,g)
 110         eor     $t2,$a,$b                       // a^b, b^c in next round
 111         eor     $t0,$t0,$T0,ror#$Sigma1[1]      // Sigma1(e)
 112         ror     $T0,$a,#$Sigma0[0]
 113         add     $h,$h,$t1                       // h+=Ch(e,f,g)
 114         eor     $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
 115         add     $h,$h,$t0                       // h+=Sigma1(e)
 116         and     $t3,$t3,$t2                     // (b^c)&=(a^b)
 117         add     $d,$d,$h                        // d+=h
 118         eor     $t3,$t3,$b                      // Maj(a,b,c)
 119         eor     $t1,$T0,$t1,ror#$Sigma0[1]      // Sigma0(a)
 120         add     $h,$h,$t3                       // h+=Maj(a,b,c)
 121         ldr     $t3,[$Ktbl],#$SZ                // *K++, $t2 in next round
 122         //add   $h,$h,$t1                       // h+=Sigma0(a)
 123 ___
 124 $code.=<<___    if ($i>=15);
 125         ror     $t0,$e,#$Sigma1[0]
 126         add     $h,$h,$t2                       // h+=K[i]
 127         ror     $T1,@X[($j+1)&15],#$sigma0[0]
 128         and     $t1,$f,$e
 129         ror     $T2,@X[($j+14)&15],#$sigma1[0]
 130         bic     $t2,$g,$e
 131         ror     $T0,$a,#$Sigma0[0]
 132         add     $h,$h,@X[$i&15]                 // h+=X[i]
 133         eor     $t0,$t0,$e,ror#$Sigma1[1]
 134         eor     $T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
 135         orr     $t1,$t1,$t2                     // Ch(e,f,g)
 136         eor     $t2,$a,$b                       // a^b, b^c in next round
 137         eor     $t0,$t0,$e,ror#$Sigma1[2]       // Sigma1(e)
 138         eor     $T0,$T0,$a,ror#$Sigma0[1]
 139         add     $h,$h,$t1                       // h+=Ch(e,f,g)
 140         and     $t3,$t3,$t2                     // (b^c)&=(a^b)
 141         eor     $T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
 142         eor     $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2]    // sigma0(X[i+1])
 143         add     $h,$h,$t0                       // h+=Sigma1(e)
 144         eor     $t3,$t3,$b                      // Maj(a,b,c)
 145         eor     $t1,$T0,$a,ror#$Sigma0[2]       // Sigma0(a)
 146         eor     $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2]   // sigma1(X[i+14])
 147         add     @X[$j],@X[$j],@X[($j+9)&15]
 148         add     $d,$d,$h                        // d+=h
 149         add     $h,$h,$t3                       // h+=Maj(a,b,c)
 150         ldr     $t3,[$Ktbl],#$SZ                // *K++, $t2 in next round
 151         add     @X[$j],@X[$j],$T1
 152         add     $h,$h,$t1                       // h+=Sigma0(a)
 153         add     @X[$j],@X[$j],$T2
 154 ___
 155         ($t2,$t3)=($t3,$t2);
 156 }
 157
 158 $code.=<<___;
 159 #include "arm_arch.h"
 160
 161 .text
 162
 163 .extern OPENSSL_armcap_P
 164 .globl  $func
 165 .type   $func,%function
 166 .align  6
 167 $func:
 168 ___
 169 $code.=<<___    if ($SZ==4);
 170         ldr     x16,.LOPENSSL_armcap_P
 171         adr     x17,.LOPENSSL_armcap_P
 172         add     x16,x16,x17
 173         ldr     w16,[x16]
 174         tst     w16,#ARMV8_SHA256
 175         b.ne    .Lv8_entry
 176 ___
 177 $code.=<<___;
 178         stp     x29,x30,[sp,#-128]!
 179         add     x29,sp,#0
 180
 181         stp     x19,x20,[sp,#16]
 182         stp     x21,x22,[sp,#32]
 183         stp     x23,x24,[sp,#48]
 184         stp     x25,x26,[sp,#64]
 185         stp     x27,x28,[sp,#80]
 186         sub     sp,sp,#4*$SZ
 187
 188         ldp     $A,$B,[$ctx]                            // load context
 189         ldp     $C,$D,[$ctx,#2*$SZ]
 190         ldp     $E,$F,[$ctx,#4*$SZ]
 191         add     $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input
 192         ldp     $G,$H,[$ctx,#6*$SZ]
 193         adr     $Ktbl,.LK$BITS
 194         stp     $ctx,$num,[x29,#96]
 195
 196 .Loop:
 197         ldp     @X[0],@X[1],[$inp],#2*$SZ
 198         ldr     $t2,[$Ktbl],#$SZ                        // *K++
 199         eor     $t3,$B,$C                               // magic seed
 200         str     $inp,[x29,#112]
 201 ___
 202 for ($i=0;$i<16;$i++)   { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
 203 $code.=".Loop_16_xx:\n";
 204 for (;$i<32;$i++)       { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
 205 $code.=<<___;
 206         cbnz    $t2,.Loop_16_xx
 207
 208         ldp     $ctx,$num,[x29,#96]
 209         ldr     $inp,[x29,#112]
 210         sub     $Ktbl,$Ktbl,#`$SZ*($rounds+1)`          // rewind
 211
 212         ldp     @X[0],@X[1],[$ctx]
 213         ldp     @X[2],@X[3],[$ctx,#2*$SZ]
 214         add     $inp,$inp,#14*$SZ                       // advance input pointer
 215         ldp     @X[4],@X[5],[$ctx,#4*$SZ]
 216         add     $A,$A,@X[0]
 217         ldp     @X[6],@X[7],[$ctx,#6*$SZ]
 218         add     $B,$B,@X[1]
 219         add     $C,$C,@X[2]
 220         add     $D,$D,@X[3]
 221         stp     $A,$B,[$ctx]
 222         add     $E,$E,@X[4]
 223         add     $F,$F,@X[5]
 224         stp     $C,$D,[$ctx,#2*$SZ]
 225         add     $G,$G,@X[6]
 226         add     $H,$H,@X[7]
 227         cmp     $inp,$num
 228         stp     $E,$F,[$ctx,#4*$SZ]
 229         stp     $G,$H,[$ctx,#6*$SZ]
 230         b.ne    .Loop
 231
 232         ldp     x19,x20,[x29,#16]
 233         add     sp,sp,#4*$SZ
 234         ldp     x21,x22,[x29,#32]
 235         ldp     x23,x24,[x29,#48]
 236         ldp     x25,x26,[x29,#64]
 237         ldp     x27,x28,[x29,#80]
 238         ldp     x29,x30,[sp],#128
 239         ret
 240 .size   $func,.-$func
 241
 242 .align  6
 243 .type   .LK$BITS,%object
 244 .LK$BITS:
 245 ___
 246 $code.=<<___ if ($SZ==8);
 247         .quad   0x428a2f98d728ae22,0x7137449123ef65cd
 248         .quad   0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
 249         .quad   0x3956c25bf348b538,0x59f111f1b605d019
 250         .quad   0x923f82a4af194f9b,0xab1c5ed5da6d8118
 251         .quad   0xd807aa98a3030242,0x12835b0145706fbe
 252         .quad   0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
 253         .quad   0x72be5d74f27b896f,0x80deb1fe3b1696b1
 254         .quad   0x9bdc06a725c71235,0xc19bf174cf692694
 255         .quad   0xe49b69c19ef14ad2,0xefbe4786384f25e3
 256         .quad   0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
 257         .quad   0x2de92c6f592b0275,0x4a7484aa6ea6e483
 258         .quad   0x5cb0a9dcbd41fbd4,0x76f988da831153b5
 259         .quad   0x983e5152ee66dfab,0xa831c66d2db43210
 260         .quad   0xb00327c898fb213f,0xbf597fc7beef0ee4
 261         .quad   0xc6e00bf33da88fc2,0xd5a79147930aa725
 262         .quad   0x06ca6351e003826f,0x142929670a0e6e70
 263         .quad   0x27b70a8546d22ffc,0x2e1b21385c26c926
 264         .quad   0x4d2c6dfc5ac42aed,0x53380d139d95b3df
 265         .quad   0x650a73548baf63de,0x766a0abb3c77b2a8
 266         .quad   0x81c2c92e47edaee6,0x92722c851482353b
 267         .quad   0xa2bfe8a14cf10364,0xa81a664bbc423001
 268         .quad   0xc24b8b70d0f89791,0xc76c51a30654be30
 269         .quad   0xd192e819d6ef5218,0xd69906245565a910
 270         .quad   0xf40e35855771202a,0x106aa07032bbd1b8
 271         .quad   0x19a4c116b8d2d0c8,0x1e376c085141ab53
 272         .quad   0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
 273         .quad   0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
 274         .quad   0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
 275         .quad   0x748f82ee5defb2fc,0x78a5636f43172f60
 276         .quad   0x84c87814a1f0ab72,0x8cc702081a6439ec
 277         .quad   0x90befffa23631e28,0xa4506cebde82bde9
 278         .quad   0xbef9a3f7b2c67915,0xc67178f2e372532b
 279         .quad   0xca273eceea26619c,0xd186b8c721c0c207
 280         .quad   0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
 281         .quad   0x06f067aa72176fba,0x0a637dc5a2c898a6
 282         .quad   0x113f9804bef90dae,0x1b710b35131c471b
 283         .quad   0x28db77f523047d84,0x32caab7b40c72493
 284         .quad   0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
 285         .quad   0x4cc5d4becb3e42b6,0x597f299cfc657e2a
 286         .quad   0x5fcb6fab3ad6faec,0x6c44198c4a475817
 287         .quad   0       // terminator
 288 ___
 289 $code.=<<___ if ($SZ==4);
 290         .long   0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
 291         .long   0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
 292         .long   0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
 293         .long   0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
 294         .long   0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
 295         .long   0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
 296         .long   0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
 297         .long   0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
 298         .long   0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
 299         .long   0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
 300         .long   0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
 301         .long   0xd192e819,0xd6990624,0xf40e3585,0x106aa070
 302         .long   0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
 303         .long   0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
 304         .long   0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
 305         .long   0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
 306         .long   0       //terminator
 307 ___
 308 $code.=<<___;
 309 .size   .LK$BITS,.-.LK$BITS
 310 .align  3
 311 .LOPENSSL_armcap_P:
 312         .quad   OPENSSL_armcap_P-.
 313 .asciz  "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
 314 .align  2
 315 ___
 316
 317 if ($SZ==4) {
 318 my $Ktbl="x3";
 319
 320 my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
 321 my @MSG=map("v$_.16b",(4..7));
 322 my ($W0,$W1)=("v16.4s","v17.4s");
 323 my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
 324
 325 $code.=<<___;
 326 .type   sha256_block_armv8,%function
 327 .align  6
 328 sha256_block_armv8:
 329 .Lv8_entry:
 330         stp             x29,x30,[sp,#-16]!
 331         add             x29,sp,#0
 332
 333         ld1.32          {$ABCD,$EFGH},[$ctx]
 334         adr             $Ktbl,.LK256
 335
 336 .Loop_hw:
 337         ld1             {@MSG[0]-@MSG[3]},[$inp],#64
 338         sub             $num,$num,#1
 339         ld1.32          {$W0},[$Ktbl],#16
 340         rev32           @MSG[0],@MSG[0]
 341         rev32           @MSG[1],@MSG[1]
 342         rev32           @MSG[2],@MSG[2]
 343         rev32           @MSG[3],@MSG[3]
 344         orr             $ABCD_SAVE,$ABCD,$ABCD          // offload
 345         orr             $EFGH_SAVE,$EFGH,$EFGH
 346 ___
 347 for($i=0;$i<12;$i++) {
 348 $code.=<<___;
 349         ld1.32          {$W1},[$Ktbl],#16
 350         add.i32         $W0,$W0,@MSG[0]
 351         sha256su0       @MSG[0],@MSG[1]
 352         orr             $abcd,$ABCD,$ABCD
 353         sha256h         $ABCD,$EFGH,$W0
 354         sha256h2        $EFGH,$abcd,$W0
 355         sha256su1       @MSG[0],@MSG[2],@MSG[3]
 356 ___
 357         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
 358 }
 359 $code.=<<___;
 360         ld1.32          {$W1},[$Ktbl],#16
 361         add.i32         $W0,$W0,@MSG[0]
 362         orr             $abcd,$ABCD,$ABCD
 363         sha256h         $ABCD,$EFGH,$W0
 364         sha256h2        $EFGH,$abcd,$W0
 365
 366         ld1.32          {$W0},[$Ktbl],#16
 367         add.i32         $W1,$W1,@MSG[1]
 368         orr             $abcd,$ABCD,$ABCD
 369         sha256h         $ABCD,$EFGH,$W1
 370         sha256h2        $EFGH,$abcd,$W1
 371
 372         ld1.32          {$W1},[$Ktbl]
 373         add.i32         $W0,$W0,@MSG[2]
 374         sub             $Ktbl,$Ktbl,#$rounds*$SZ-16     // rewind
 375         orr             $abcd,$ABCD,$ABCD
 376         sha256h         $ABCD,$EFGH,$W0
 377         sha256h2        $EFGH,$abcd,$W0
 378
 379         add.i32         $W1,$W1,@MSG[3]
 380         orr             $abcd,$ABCD,$ABCD
 381         sha256h         $ABCD,$EFGH,$W1
 382         sha256h2        $EFGH,$abcd,$W1
 383
 384         add.i32         $ABCD,$ABCD,$ABCD_SAVE
 385         add.i32         $EFGH,$EFGH,$EFGH_SAVE
 386
 387         cbnz            $num,.Loop_hw
 388
 389         st1.32          {$ABCD,$EFGH},[$ctx]
 390
 391         ldr             x29,[sp],#16
 392         ret
 393 .size   sha256_block_armv8,.-sha256_block_armv8
 394 ___
 395 }
 396
 397 $code.=<<___;
 398 .comm   OPENSSL_armcap_P,4,4
 399 ___
 400
 401 {   my  %opcode = (
 402         "sha256h"       => 0x5e004000,  "sha256h2"      => 0x5e005000,
 403         "sha256su0"     => 0x5e282800,  "sha256su1"     => 0x5e006000   );
 404
 405     sub unsha256 {
 406         my ($mnemonic,$arg)=@_;
 407
 408         $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
 409         &&
 410         sprintf ".inst\t0x%08x\t//%s %s",
 411                         $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
 412                         $mnemonic,$arg;
 413     }
 414 }
 415
 416 foreach(split("\n",$code)) {
 417
 418         s/\`([^\`]*)\`/eval($1)/geo;
 419
 420         s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo;
 421
 422         s/\.\w?32\b//o          and s/\.16b/\.4s/go;
 423         m/(ld|st)1[^\[]+\[0\]/o and s/\.4s/\.s/go;
 424
 425         print $_,"\n";
 426 }
 427
 428 close STDOUT;